[
https://issues.apache.org/jira/browse/SPARK-9347?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14641763#comment-14641763
]
Samphel Norden commented on SPARK-9347:
---------------------------------------
Consider a top level folder with folder hierachy as follows
root/parquet_date=20150717/parquet_hour_of_day=00
For each date there are 24 hour folders
Overall there are approx 400 parquet (snappy compressed) files per hour folder
and we are currently looking at 8 days worth of data
So roughly 400*24*8 days of files
After launching spark shell, doing this almost always hangs
scala> sqlContext.parquetFile(<top level> folder)
It appears to be directly proportional to the number of files in the folder and
it appears that the metadata load is reading every single file.
Verified by doing above against a folder with order of magnitude less files
returns quickly.
Is there a better way to load the data into a dataframe.
> spark load of existing parquet files extremely slow if large number of files
> ----------------------------------------------------------------------------
>
> Key: SPARK-9347
> URL: https://issues.apache.org/jira/browse/SPARK-9347
> Project: Spark
> Issue Type: Improvement
> Components: SQL
> Affects Versions: 1.3.1
> Reporter: Samphel Norden
>
> When spark sql shell is launched and we point it to a folder containing a
> large number of parquet files, the sqlContext.parquetFile() command takes a
> very long time to load the tables.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
